Title: Similarity analysis and modeling in mobile societies : the missing link
Full Citation
Permanent Link: http://ufdc.ufl.edu/UF00101377/00001
 Material Information
Title: Similarity analysis and modeling in mobile societies : the missing link
Physical Description: Book
Language: English
Creator: Thakur, Gautam
Publisher: Department of Computer and Information Science and Engineering, University of Florida
Place of Publication: Gainesville, Fla.
Copyright Date: 2010
Abstract: A new generation of 'behavior-aware' delay tolerant networks is emerging in what may define future mobile social networks. With the introduction of novel behavior-aware protocols, services and architectures, there is a pressing need to understand and realistically model mobile users behavioral characteristics, their similarity and clustering. Such models are essential for the analysis, performance evaluation, and simulation of future DTNs. This paper addresses issues related to mobile user similarity, its definition, analysis and modeling. To define similarity, we adopt a behavioral-profile based on users location preferences using their on-line association matrix and its SVD,then calculate the behavioral distance to capture user similarity.This measures the difference of the major spatio-temporal behavioral trends and can be used to cluster users into similarity groups or communities. We then analyze and contrast similarity distributions of mobile user populations in two settings: (i) based on real measurements from four major campuses with over ten thousand users for a month, and (ii) based on existing mobility models, including random direction and time-varying community models. Our results show a rich set of similar communities in real mobile societies with distinct behavioral clusters of users. This is true for all the traces studied, with the trend being consistent overtime. Surprisingly, however, we find that the existing mobility models do not explicitly capture similarity and result inhomogeneous users that are all similar to each other. Thus the richness and diversity of user behavioral patterns is not captured to any degree in the existing models. These findings strongly suggest that similarity should be explicitly captured in future mobility models, which motivates the need to re-visit mobility modeling to incorporate accurate behavioral models in the future
 Record Information
Bibliographic ID: UF00101377
Volume ID: VID00001
Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.


This item has the following downloads:

PDF ( 2 MBs ) ( PDF )

Full Text

Similarity Analysis and Modeling

in Mobile Societies: The Missing Link

Gautam S. Thakur*, Ahmed Helmy*, Wei-Jen Hsu+
*Computer and Information Science and Engineering Department, University of Florida, Gainesville, FL
+Cisco Systems Inc., San Jose, CA
{gsthakur,helmy}@cise.ufl.edu, wehsu@cisco.com

A new generation of "behavior-aware" delay tolerant
networks is emerging in what may define future mobile social
networks. With the introduction of novel behavior-aware
protocols, services and architectures, there is a pressing need to
understand and realistically model mobile users behavioral
characteristics, their similarity and clustering. Such models are
essential for the analysis, performance evaluation, and simulation
of future DTNs. This paper addresses issues related to mobile user
similarity, its definition, analysis and modeling. To define
similarity, we adopt a behavioral-profile based on users location
preferences using their on-line association matrix and its SVD,
then calculate the behavioral distance to capture user similarity.
This measures the difference of the major spatio-temporal
behavioral trends and can be used to cluster users into similarity
groups or communities.
We then analyze and contrast similarity distributions of
mobile user populations in two settings: (i) based on real
measurements from four major campuses with over ten thousand
users for a month, and (ii) based on existing mobility models,
including random direction and time-varying community models.
Our results show a rich set of similar communities in real
mobile societies with distinct behavioral clusters of users. This is
true for all the traces studied, with the trend being consistent over
time. Surprisingly, however, we find that the existing mobility
models do not explicitly capture similarity and result in
homogeneous users that are all similar to each other. Thus the
richness and diversity of user behavioral patterns is not captured
to any degree in the existing models. These findings strongly
suggest that similarity should be explicitly captured in future
mobility models, which motivates the need to re-visit mobility
modeling to incorporate accurate behavioral models in the future.

Behavior-aware DTNs, trace analysis, similarity, clustering,
mobility modeling.

1. Introduction
Future social networks are expected to have classes of
applications that are aware of mobile users' behavioral profiles
and preferences and are likely to support peer-to-peer mobile
networking including delay tolerant networks (DTNs). A new
generation of protocols is emerging, including behavior-aware
communication paradigms (such as profile-cast [15]) and service
architectures (such as participatory sensing [35, 43]).
Such behavior-aware communication paradigm leverages
user behavior and preferences to achieve efficient operation in
DTNs (e.g., interest-based target message forwarding; encounter-
based routing, mobile resource discovery). Accurate models of
mobile user behavioral profiles are essential for the analysis,

performance evaluation, and simulation of such networking
Hence, there is a compelling need to understand and
realistically model mobile users behavioral profiles, similarity and
clustering of user groups.
Earlier work on mobility modeling presented advances in
random mobility models (e.g., RWP, RD [5]), synthetic models
that attempt to capture spatial correlation between nodes (e.g.,
group models [2]) or temporal correlation and geographic
restrictions (e.g., Freeway, Manhattan, Pathway Models [1]).
More recent models tend to be trace-driven and some account for
location preferences and temporal repetition [19]. However,
similarity characteristics between clusters of nodes, which lie in
the heart of behavior-aware networking, have not been modeled
explicitly by these mobility models. Hence, it is unclear whether
(and to which degree) similarity is captured.
In this paper, we address issues related to mobile user
similarity, its definition, analysis and modeling. Similarity, in this
study, is defined by mobility preferences, and is meant to reflect
the users interests to the extent that can be captured by wireless
measurements of on-line usage. To define similarity, we adopt a
behavioral-profile based on users mobility and location
preferences using an on-line association matrix representation,
then use the cosine product of their weighted Eigen-behaviors to
capture similarity between users. This quantitatively compares the
major spatio-temporal behavioral trends between mobile network
users, and can be used for clustering users into similarity groups
or communities. Note that this may not reflect social ties between
users or relationships per se, but does reflect mobility-related
behavior that will affect connectivity and network topology
dynamics in a DTN setting.
We analyze similarity distributions of mobile user
populations in two settings. The first analysis aims to establish
deep understanding of realistic similarity distributions in such
mobile societies. It is based on real measurements of over 8860
users for a month in four major university campuses, USC [38]
MIT, Dartmouth [27] and UF. It may be reasonable to expect
some clustering of users that belong to similar affiliations, but
quantification of such clustering and its stability over time is
necessary for developing accurate similarity models. Furthermore,
on-line behavior that reflects distribution of active wireless
devices may not necessarily reflect work or study affiliations or
social clustering. For DTNs, on-line activity and mobility
preferences translate into encounters that are used for
opportunistic message forwarding, and this is the focus of our
study rather than social relations per se.
The second similarity analysis we conduct aims to
investigate whether existing mobility models provide a reasonable
approximation of realistic similarity distributions found in the
campus traces. It is based on existing mobility models where we

analyze a commonly used random-based mobility model (the
random direction (RD) model [5]) and an advanced trace-driven
mobility model; (the time-variant community (TVC) model [19]).
Our results show that among mobile users, we can discover
distinct clusters of users that are similar to each other, while
dissimilar to other clusters. This is true for all campuses, with the
trend being consistent and stable over time. We find an average
modularity of 0.64, clustering coefficient of 0.86 and path length
of 0.24 among discovered clusters. Surprisingly, however, we find
that the existing mobility models do not explicitly capture
similarity and result in homogeneous users that are all similar to
each other (in one big cluster). This finding generalizes to all
other mobility models that produce homogeneous users, not only
the mobility models studied in this paper. Thus the richness and
diversity of user behavioral patterns is not captured in any degree
in the existing models.
Our findings strongly suggest that unless similarity is
explicitly captured in mobility models, the resulting behavioral
patterns are likely to deviate dramatically from reality, sometimes
totally missing the richness in the similarity distribution found in
the traces. Furthermore, this indicates our current inability to
accurately simulate and evaluate similarity-based protocols,
services and architectures using mobility models. This motivates
the need to re-visit mobility modeling to incorporate accurate
behavioral models in the future. Our analysis is the first to provide
insights leading to that direction.

2. Related Work
A new paradigm of protocols that relies on the human
behavioral patterns has gained recent attention in DTN-related
research. In these studies, researchers attempted to use social
aspects of human mobility to derive new services and protocols.
The study in [8] has used social network metrics for information
flow and behavioral routing in DTNs and shown that using nodes
with high betweenness as forwarders achieves best overall
delivery performance. A similar study in [6, 20, 21] has used
various centrality measures and proposed an array of forwarding
algorithms for information dissemination in opportunistic
networks. A new behavior-oriented service, called profile-cast,
that relies on tight user-network integration was introduced in
[15]. Profile-cast provides a systematic framework to utilize
implicit relationships discovered among mobile users for interest-
based message efficient forwarding and delivery in DTNs.
Participatory sensing [35, 39, 43] provides a service for crowd
sourcing using recruiting campaigns using mobile user profiles
[16]. All these works rely on and utilize similarity of mobile user
Modeling and simulating profile similarity accurately in
networked mobile societies is thus imperative to the design and
evaluation of these classes of networking protocols and services.
The aim of this paper is to identify the underlying similarity
structure that govern mobile societies and provide measures of its
existence in common mobility models both quantitative and
There is a large body of work on mobility modeling and trace
analysis, but we shall only discuss briefly (for lack of space) some
of the main approaches. Generally, mobility models are either
synthetic or trace-driven. The most common models include
random mobility generators, including random direction, random
waypoint and random walk. In addition, several other models
attempt to capture spatial correlation (e.g., group mobility),
temporal correlation (e.g., freeway) and geographic restrictions

(e.g., pathway, manhattan) [1]. These models are useful for simple
initial evaluations, but not based on real traces and have been
shown to lack important characteristics of user location
preferences and periodicity [19]. Some community models [12,
34] represent social connections based on social network theory,
but do not capture mobility related preferences of similarity.
Trace-based mobility models, by contrast, derive their parameters
from analysis of measurements and can reproduce user
preferences [19]. There has been no prior evaluation, however, of
similarity and clustering in mobility models. Recently, several
studies have attempted to analyze realistic behavioral through
processing of WLAN traces. In [26, 47] individual behavior
(preferences) of users was studied, while in [17] nodal encounters
have been analyzed. In [16] users preferences are captured
through an association matrix and analyzed. We utilize a similar
representation here. This study, however, is the first to compare
trace-based similarity clustering to model-based similarity
clustering and show the need to re-visit mobility modeling to
capture similarity characteristics.

3. Similarity
The congregation of mobile agents with similar characteristic
patterns naturally develop mobile societies in wireless networks
[10, 22, 48]. Upon reflection it should come as no surprise that
these characteristics in particular also have a big impact on the
overall behavior of the system [7, 21, 32, 33]. Researchers have
long been working to infer these characteristics and ways to
measure them. One major observation is that people demonstrate
periodic reappearances at certain locations [9, 16, 24], which in
turn breeds connection among similar instances [31]. Thus, people
with similar behavioral principle tie together. This brings an
important aspect where, user-location coupling can be used to
identify similarity patterns in mobile users. So, for the purpose of
our study, to quantify similarity characteristics among mobile
agents, we use their spatio-temporal preferences and preferential
attachment to locations and the frequency and duration of visiting
these locations.
It is important to study similarity in DTN to develop
behavioral space for efficient message dissemination [15] and
design behavior-aware trust advisors among others [28]. For
efficient networking, it can help to quantify traffic patterns and
develop new protocols and application to target social networking.
Analysis of similarity can be used to evaluate the network
transitivity, which helps to analyze macro-mobility, evolutionary
characteristics and emergent properties.
In this section, we introduce association matrix that captures
spatio-temporal preferences and a statistical technique that use it
to measures similarity among mobile users.

3.1 Capturing Spatio-Temporal Preferences
We use longitudinal wireless activity session to build mobile
user's spatio-temporal profile. An anonymous sample is shown in
Table 1. Each entry of this measurement trace has the location of
association and session time information for that user. The
location association coupled with time dimension provides a good
estimate of user online mobile activity and its physical proximity
with respect to other online users [17, 26]. We devise a scalable
representation of this information in form of an association matrix
as shown in Figure 1. Each individual column corresponds to a
unique location in the trace. Each row is an n-element association
vector, where each entry in the vector represents the fraction of
online time the mobile user spent at that location, during a certain

Table 1: Anonymized Sample of Mobile user WLAN session

Node Mac ID Location Start Time End Time
aa:bb:cc:dd:ee:ff Loc-1 64400343 66404567
aa:bb:cc:dd:ee:ff Loc-2 85895623 86895742

aa:bb:cc:dd:ee:ff Loc-3 87444343 89404567

aa:bb:cc:dd:ee:ff Loc-4 98846767 99878766

time period (which can be flexibly chosen, such as an hour, a day,
etc.). Thus for n distinct locations and t time periods, we generate
a t-by-n size association matrix.
Representation Flexibility: The representation of spatio-temporal
preferences in form of an association matrix can be changed to
use each column for a building (where a collection of access
points represent a building) and the time granularity can be
changed to represent hourly, weekly or monthly behavior. For the
purpose of our study, each row represents a day in the trace and
column represents an individual access point.

3.2 Characterizing Association Patterns
For a succinct measure of mobile user behavior, we capture
the dominant behavioral patterns by using Singular Value
Decomposition (SVD) [14] of the association matrix. SVD has
several advantages: 1) It helps to convert high dimensional and
high variable data set to lower dimensional space there by
exposing the internal structure of the original data more clearly. 2)
It is robust to noisy data and outliers. 3) It can easily be
programmed for handheld devices, which is our other on-going
The Singular Value Decomposition of a given matrix A can
be represented as a product of three matrices: an orthogonal
matrix U, a diagonal matrix S, and the transpose of an orthogonal
matrix V. It is written as:
A = U- S VT
Where, UT. U=I=VTV, U is t-by-t matrix whose columns
are orthonormal eigenvectors of AAT, S is a t-by-n matrix with r
non-zero entries on its main diagonal containing the square roots
of eigen values of matrix A in descending order of magnitude and
VT is a n-by-n matrix whose columns are the orthonormal
eigenvectors of ATA. Thus the eigen behavior vectors of
V={v,,v2,3,...v } summarize the important trends in the
original matrix A. The singular values of S = {s,s2,S3,...,sr}
ordered by their magnitude {S1 2 S2 2 SJr. The percentage
of power captured by each eigen vector of the matrix A is
calculated by



It has been shown that [16] SVD achieves great data reduction
on the original association matrix and 90% or more power for
most of the users is captured by five components of the
association vectors. By this result, we infer that user's few top
location-visiting preferences are more dominant than the
remaining ones.

Each row represents the
percentage of time spent at each / Dorm Office
location for a day
"1. I 0,4 - 0 I
An entry represents the s I
percentage of online time during c -j
time day iat location j

Each column corresponds to a location

0 0 0.5 0.3 0.2
0.2 0.3 0.4 0.1 0
A= 0 0 0 0.6 0.4
0.2 0.3 0.1 0.2 0.2
0.1 0 0 0 0.9

Figure 1: (a) A prototype ofAssociation Matrix. The columns represent
locations (access point, building, etc) and rows represent time
granularity (days, weeks, etc.). (b) A computed matrix A with 5
locations and time periods. Each entry represent the percentage online
time spent at corresponding location column.

3.3 Calculating Similarity
We use the eigen vectors of association matrix A to quantitatively
measure the similarity between behavioral profiles of mobile user
pairs. For a pair of users, with respective eigen-vectors as
X ={x1,x2,x3,...x } and Y= f{yl,y23,...y }, the behavior
similarity can be calculated by the weighted sum of pair wise
inner product of their eigen vectors as
Sim(X,Y) = W wYj xi ,- y
i=1 j=1
Sim(X,Y) is quantitative measure index that shows the closeness
of two users in spatio-temporal dimension. The value of similarity
lies between 0 s Sim(X,Y) 5 1. A higher value is derived from
users with similar association patterns. In this study, we are the
first one to investigate the distribution of such a similarity metric
among user pairs based on realistic data sets.

4. Trace-Based Similarity Analysis
In this section, we obtain the distribution of Similarity among
mobile users and apply a divisive quality function -modularity
[36] to discover mobile societies in wireless networks. We expect
to see a natural division of mobile nodes into densely connected
clusters. Each of these clusters consists of users with similar
spatio-temporal preferences. For the purpose of this study, we
examine very large real world wireless data measurements of four
university campuses collected for a period of several months with
thousands of users.
4.1 Dataset and Trace Analysis
WLAN dataset from four university campuses are considered
as shown in Table 2. We collect these datasets from the publicly
available MobiLib[38] and Crawdad[27] repositories. Table 2
provides the detail of these WLAN measurements. We chose
university campuses because they are extensive, have high density


Table 2: Details of Wireless Measurements

Campus # Users Duration
Dartmouth 1500 Fall 2007
MIT 1366 Fall 2006
Univ. of Florida 3000 Fall 2008
USC 3000 Fall 2007

of active users and include location information. Also, these
datasets have been used in previous studies of mobility modeling
[19, 21, 24, 48]. We perform Systematic Random Sampling [42]
on the datasets to get an unbiased subset of mobile users from the
population. Table 2 specifies the sampling frame that we use for
this study. In the second step, we extract relevant statistics of
mobile user spatio-temporal patterns. In the third step, for each
mobile user we obtain normalized association matrix as shown in
Figure 1 with time granularity of one day. On this matrix we
apply SVD to extract the dominant trends. Finally, we compute

Dartmouth Campus

Similarity Score

UF Campus
+ 1 week
-*- 2 weeks
+ : --+- 3 weeks
L- 4 weeks

+ q


.- 4-
: '*1^;:- *~ ^

'*: l ^ ^ . .

0 0.2 0.4 0.6
Similarity Score

0.8 1

4.2 Similarity Analysis
The distribution histogram of similarity scores for the
campus datasets is shown in Figure 2. The figure shows number of
user pairs as a function of similarity score that quantify the
behavioral similarity between mobile users. We observe that: 1)
mobile societies compose of users with mixed behavioral
similarities, 2) For all four time periods there is a consistency and
stability in the similarity score among mobile user pairs. The low
similarity scores (0 0.1) in Figure 2 indicate a substantial portion
of users is spatio-temporally very dissimilar. On the other hand,
similarity scores of (0.9 -1.0) suggest a statistically significant
likelihood of high-density ties creating tightly knit groups. The
variation in the middle shows partially similar and partially
dissimilar user pairs. This is significant and provides an insight
into the existence of mobile societies in the network with quite
similar location visiting preferences. Overall, the curves show an
assortative mixing of user pairs for all possible similarity scores.
Figure 3 gives a normalized log plot to compare data sets from

MIT Campus

Similarity Score


Similarity Score

(C) (d)
Figure 2: Similarity distribution histogram among user pairs is shown. All the four time intervals show near consistent user pair counts for a particular
similarity score. Lowest similarity score (0.0 0.1) shows that users have very different spatio-temporal preferences. A fraction of the user pairs are also
very similar with (0.9 1.0) similarity score.

the cosine similarity of all user pairs. We perform this process
iteratively for four different time intervals: 1 week, 2 weeks, 3
weeks and 4 weeks.

different campuses, and shows that similarity exists evenly across
all the traces. Next, we briefly explain modularity and use a
divisive algorithm to discover mobile societies in the traces.

Table 3: Network Analysis of Datasets on three different metrics

Dataset Clustering Average Path Modularity
Coefficient Length
Ori Rand Ori Rand Ori Rand
Dartmouth 0.89 0.05 0.10 2.47 0.63 0.2
MIT 0.92 0.05 0.40 2.12 0.79 0.14
UF 0.78 0.051 0.30 2.605 0.67 0.24
USC 0.91 .05 0.19 2.0 0.46 0.11
*Ori = Original Dataset Graph *Rand = Random Graph

4.2.1 Modularity
To understand the underlying structure of mobile societies
(or communities), the similarity distribution is not sufficient.
Therefore, we use a robust method to segregate user pairs that
have high similarity score into tightly knit groups. To detect such
communities in a graph like structure, a centrality-index-driven
method [13] is utilized. This measure to detect communities
circumvents the traditional clustering notion to identify most
central edges. Instead, a divisive algorithm is applied based on
identifying least central edges, which connect most communities
(via edge betweenness). First, the betweenness score of edges are
calculated as the number of shortest paths between pair of vertices
that run through it. Understandably, tightly knit communities are
loosely connected by only few intergroup edges and hence
shortest paths traverse these edges repeatedly, thereby increasing
their respective betweenness score. If such edges are removed,
according to a threshold, what we get are the groups of tightly
knitted vertices known as communities. To identify a reasonable
threshold value, modularity is used. Modularity is the difference
of edges falling within communities and the expected number in
an equivalent network with randomly-placed edges [13, 36, 37].
4.2.2 Detection ofMobile Societies
Human networks are known to exhibit a multitude of
emergent properties that characterize the collective dynamics of a
complex system [23, 45, 49]. Their ability to naturally evolve into
groups and communities is the reason they show non-trivial
clustering. Here, we consider the spatio-temporal preferences and
cosine similarity of mobile users as a relative index to generate
emergent structures, which we call mobile societies. The network
transitivity structures of mobile nodes for various campus datasets
are shown in Figure 4. We use mutual similarity score of mobile
nodes to produce a connected graph and applied random iterations
of modularity [3] and betweenness algorithm to infer the mobile
societies. A set of visibly segregated clusters validates their
detection and presence in mobile networks. Modularity Analysis for Mobile Societies
Statistically, modularity greater than 0.4 is considered
meaningful in detecting community structure. For our dataset, we
also find high modularity index as compared to an equivalent
random graph. The comparison is shown in Table 3. Henceforth,
the heterogeneity in dataset has tightly knitted Mobile Societies.
This analysis further helped us to investigate the possibility of
existence of different clusters of users based on their proximity in
similarity score values.




' -0-MIT
--*- USC
S- -Dartmouth



0.1 0.2 0.3 0.4 0.5 0.6 0.7
Similarity Score

O- 0

0.8 0.9 1

Figure 3: Log Normalized Similarity distribution of all four data sets is
shown. Network Analysis for Mobile Societies
We compute the average clustering coefficient and the mean-
shortest path length of these clusters. We compare the results with
a random graph of the same size to understand the variation and
capacity to depict small world characteristics. Table 3, delineates
network properties and average modularity that provide details of
the structure of mobile societies against same size random graph.
The comparative values in the table clearly show that mobile
societies can exhibit small world characteristics. However, we
leave such small world study for future work.
Based on the above analysis, we find that similarity not only
exists among mobile users, but its distributions seem to be stable
for different time periods. Furthermore, this trend is consistent in
all four traces, which highlights similarity clustering as an
important characteristic to capture using mobility models.

5. Similarity in Models The Missing Link
In this section, we evaluate existing mobility models and
contrast their output against real trace results. Trace based
mobility models [4, 6, 11, 18, 25, 29, 30, 40] are a close
approximation of realistic human movements and their non-
homogenous behavior. They focus on vital mobility properties
like node's on/off behavior, connectivity patterns, spatial
preferences under geographical restrictions, contact duration,
inter-meeting and pause time, etc. We consider two mobility
models, the random direction model (a widely used "classic"
mobility model) and Time Variant Community Model [19] (due
to its capability to capture spatio-temporal mobility properties). In
the ensuing text, we briefly describe the TVC model and use it to
generate realistic movements. Finally, we compare its result
against the similarity characteristic found in real measurements.
5.1 TVC Model
The TVC model [19] is proposed to capture two prominent
features in wireless network user mobility observed in real traces
(1) skewed location visiting preferences and (2) periodical re-
appearance at the same location. TVC model introduces multiple
"preferred geographical locations", or the "communities", to
which the mobile nodes visit often, in order to capture the fact that
most mobile users spend significant portion of time at a few
locations. Further, the TVC model also introduces a structure in

time (the "time periods") that allows setting up different mobility
preferences for users in a periodic fashion. TVC model is the best
we have found in literature in terms of the capacity to closely
reproduce realistic user behavior and the flexibility to be fine-
tuned for different environments. Secondly, TVC also encompass
properties depicted by [41, 44], which implicitly helps us to
evaluate similarity on these models also. Thus, we adopt the TVC

curves for both campuses. We clearly observe a discrepancy
between the curves from actual traces and the two mobility
models (TVC and random direction). In addition, dendrograms in
Figure 6 shows the result of hierarchical clustering based on user's
mutual similarity scores. Here, in real traces we find clusters at
different similarity scores. In Figure 6(a), the average distance of
2.0 has close to 18 small clusters and Figure 6(c) shows 16 small

(a) Dartmouth Campus (b) MIT Campus (c) UF Campus (d) USC Campus

Figure 4: Shown are the structural and spatio-temporal dynamics of Mobile Societies as function of weighted cosine similarity score, produced from highly
positive modularity values. Note: this figure is best viewed in color.

model in this study and evaluate its capacity to capture the social
structure we observed in section 4.

5.2 TVC Model Evaluation
We setup the TVC model for two university campuses (MIT
and USC) to statistically evaluate the similarity metric established
previously. Our goal is two folds:
1. As proposed by the TVC model, we seek to maintain
the skewed location visiting preferences and time
dependent mobility behavior of users.
2. To analyze whether TVC model successfully captures
similarity among mobile users and quantitatively
simulate the distribution that we have seen in the real

5.2.1 Construction of TVCModel for Campuses
Initially, we determine the number of communities that
nodes should periodically visit. We determine that top 2-3
communities capture most skewed location visiting preferences.
Then we employ a weekly time schedule to capture the periodic
re-visits to these major communities. To keep fair comparison
against the real measurements, we configure the TVC model with
same number of mobile nodes and generating measurements
equivalent to one month time period with one-day granularity.
Finally, for WLAN measurement we assume mobile users are
stationary while being online [19].

5.2.2 Similarity Evaluation
TVC model accurately demonstrates location visiting
preferences and periodic reappearances for both campuses [46].
Surprisingly, it is unable to accurately capture the richness in
similarity distribution on spatio-temporal basis. For all values of
similarity score except 0.9, TVC and Random Direction model
yields no user pairs. Figure 5 shows similarity distribution CDF

clusters of mobile users. However, corresponding TVC

. ......-. ....I-

*0 .4"

04 MIT-Trace 34-weeks
r ---MIT-Trace -weeks
-" "MIT-Trace 2-weeks
S- nMIT-Trace 1-week
"" -MnT-Model[TVC] /
X MIT-Rimdom Direction

S 0 02 03 06 07 08 1
Similarity Score

'------------ ---------------*-*-------
4 ----------.- -.----------
S- I
? .- -- ~-- *- -- -;

S* USC-Trace 4-week-
4 -I 1 SC-Trace 3-weeks
o --- 1 SC-Trace 2-weeks
S-* *USC-Tracc 1-weck
[ "USC-Modcl[T'VC] I
x USC-Random Direction

S ol o 0o 01 Oj as 07 o8 o9
Similarity Score

Figure 5:Cumulative distribution function of distances for the similarity
score of mobile users. Real trace curves show a conformance with user
pairs for different values of similarity score, while TVC and Random
Direction Model has all users pairs in the 0.9 score range.

"' .

;c: -.
ii r: :

.;, ..
,... ...

.Y- ".. .

- ,. i: 1...,

. 1


.. .
. .;



r 4..

MIT Campus Clustering

Mobile Users

Mobile Users

72913 2 62126 42427201228 51718 110 923112515192216 314 830
Mobile Users

USC TVC Clustering

Mobile Users

Figure 6: Dendrograms giving visual representation of two-dimensional hierarchical clustering for real and TVC model generated traces for USC and
MIT campus mobile users. Real traces (Figure a & c) show an incremental built-up of component based on the similarity score strength between mobile
user. TVC Model (Figure b & d), output only one cluster containing all mobile users. Invariably, TVC treats all mobile users to have same preferences.

dendrograms in Figure 6(b) and 6(d) show only one cluster of
mobile users at a distance of 2.0. A possible explanation is that the
community assignment in TVC model creates a homogeneous
user population where all nodes are assigned the same
communities. While it captures the location visiting and periodic
preferences, it fails to differentiate among mobile nodes with
different behaviors. What is missing here is a mechanism to
assign different locations as the communities to different nodes, in
a way that reproduces the social structure (clusters) observed in
the traces.
Results in this section show that although TVC model is able
to capture location visiting preferences and periodic
reappearances, it does not capture the similarity metric
distribution and the clusters with different behaviors in the traces.
Random direction model also fails on this front in a similar way.
This study realizes us that current mobility models are not fully
equipped to handle behavioral metrics and community behavior of
users that form mobile societies. It compels us to revisit mobility
modeling in the attempt to capture both individual and community
behavior of mobile users, which is part of our on-going study.

6. Conclusion and Future Work
Several novel behavior-aware protocols and services are
being designed for DTNs, using similarity as a corner stone for
their architecture. In this paper, we analyze the spatio-temporal
behavioral similarity profiles among mobile users. We define
mobility profiles based on users association matrices, then use a
SVD-based-weighted-cosine similarity index to quantitatively
compare these mobility profiles. Analysis of extensive WLAN
traces from four major campuses reveals rich similarity

distribution histograms suggesting a clustered underlying
structure. Application of modularity based clustering validated
and further quantified the clustered behavior in mobile societies.
Similarity graphs exhibit an average modularity of 0.64, and
clustering coefficient of 0.86, which indicates potential for further
small world analysis. Finally, we compared similarity
characteristics of the traces to those from existing common and
community based mobility models to capture similarity.
Surprisingly, existing models are found to generate a
homogeneous community with one cluster and thus deviate
dramatically from realistic similarity structures. This indicates a
serious flaw in existing models and their lack of support for
emerging behavior-aware protocols, and provides a compelling
motivation to re-visit mobility modeling. We believe this is vital
for the evaluation and design of next-generation behavior-aware
protocols likely to be used in DTN and other ad hoc networks.
In the future, we plan to further investigate similarity
modeling and its effect on protocol efficiency and routing
decisions in mobile networks. We also plan to perform stability,
evolution and sensitivity analysis of similarity across different
time and space granularities. Furthermore, we shall introduce a
new mobility model with a systematic procedure to capture
realistic similarity structures in mobile societies. The key insight,
based on this study, is to assign communal probabilities to nodes
using explicit parameters to capture inter-dependencies between
behavioral similarity clusters.

MITTVC Clustering


[1] F. Bai, and Helmy, A., A Survey of Mobility Modeling and Analysis in
Wireless Adhoc Networks, Book Chapter, Oct. 2006.
[2] Bai, F., Sadagopan, N. and Helmy, A. IMPORTANT: A framework to
systematically analyze the Impact of Mobility on Performance of
RouTing protocols for Adhoc NeTworks. IEEE INFOCOM, 2003.
[3] Bastian, M., Heymann, S. and Jacomy, M. Gephi: An Open Source
Software for Exploring and Manipulating Networks. International AAAI
Conference on Weblogs and Social Media, 2009.
[4] Bhattacharjee, S. D., Rao, A., Shah, C., Shah, M. and Helmy, F. A.
Empirical modeling of campus-wide pedestrian mobility, IEEE
Vehicular Technology Conf(VTC), 2004.
[5] Camp, T., Boleng, J. and Davies, V. A survey of mobility models for
ad hoc network research. Wireless Communications and Mobile
Computing, pp. 483-502, 2002.
[6] Chaintreau, A., Hui, P., Crowcroft, J., Diot, C., Gass, R. and Scott, J.
Impact of Human Mobility on Opportunistic Forwarding Algorithms.
IEEE Transactions on Mobile Computing, pp. 606-620, 2007.
[7] Costa, P., Mascolo, C., Musolesi, M. and Picco, G. P. Socially-aware
Routing for Publish-Subscribe in Delay-tolerant Mobile Ad Hoc
Networks. IEEE Jml on Selected Areas in Communications, June 2008.
[8] Daly, E. M. and Haahr, M. Social Network Analysis for Information
Flow in Disconnected Delay-Tolerant MANETs. IEEE Transactions on
Mobile Computing, pp. 606-621, 2009.
[9] Eagle, N. and Pentland, A. Eigenbehaviors: Identifying Structure in
Routine. Proc. Roy. Soc. A, 2006.
[10] Eagle, N., Pentland, A. and Lazer, D. Inferring friendship network
structure by using mobile phone data. Proc. Of the National Academy of
Sciences, 106, 36, September 8 2009.
[11] Ekman, F., Ker\anen, A., Karvo, J. and Ott, J. Working day movement
model. ACM workshop on Mobility models, pp. 33-40, 2008,.
[12] Garbinato, B., Miranda, H. and Rodrigues, L. Middleware for Network
Eccentric and Mobile Applications. Springer Publishing Company,
Incorporated, 2009.
[13] Girvan, M. and Newman, M. E. J. Community Structure in Social and
Biological Networks. Proc. Natl. Acad. Sci., pp. 7821-7826Jun 2002.
[14] Horn, R. A. and Johnson, C. R. Matrix Analysis. Cambridge University
Press, 1990.
[15] Hsu, W., Dutta, D. and Helmy, A. CSI: A Paradigm for Behavior-
oriented Delivery Services in Mobile Human Networks, IEEE/ACM
Transactions on Networking (submitted), 2009.
[16] Hsu, W., Dutta, D. and Helmy, A. Mining behavioral groups in large
wireless LANs. ACMMobiCom, pp. 338-341, 2007.
[17] Hsu, W. and Helmy, A. On nodal encounter patterns in wireless LAN
traces. IEEE Int.l Workshop on Wireless Network Measurements, 2006.
[18] Hsu, W., Merchant, K., Shu, H., Hsu, C. and Helmy, A. Weighted
waypoint mobility model and its impact on ad hoc networks. ACM
SIGMOBILE Mob.Comp.Commun.Rev, pp. 59-63, 2005.
[19] Hsu, W., Spyropoulos, T., Psounis, K. and Helmy, A. Modeling spatial
and temporal dependencies of user mobility in wireless mobile networks.
IEEE/ACM Transactions on Networking, 17, 5, pp. 1564-1577, 2009.
[20] Hui, P. and Crowcroft, J. Human mobility models and opportunistic
communications system design. Philosophical Transactions of the Royal
Society A: Mathematical, Physical and Engineering Sciences, June 2008.
[21] Hui, P., Crowcroft, J. and Yoneki, E. Bubble rap: social-based
forwarding in delay tolerant networks. ACMMobiHoc,pp. 241-250, 2008.
[22] Hui, P., Yoneki, E., Chan, S. Y. and Crowcroft, J. Distributed
community detection in delay tolerant networks. ACM MobiArch
workshop on Mobility in the evolving internet architecture, 2007.
[23] J., C., M.C., G., P., W., T., S., G., M. and A.-L., B. Uncovering
individual and collective human dynamics from mobile phone records.
Journal of Physics A Mathematical General, 41, 22, Jun 2008.
[24] Kim, M. and Kotz, D. Periodic properties of user mobility and access-
point popularity. Personal Ubiquitous Comput., pp. 465-479, 2007.
[25] Kim, M., Kotz, D. and Kim, S. Extracting a mobility model from real
user traces. IEEE INFOCOM, April, 2006.
[26] Kotz, D. and Essien, K. Analysis of a Campus-wide Wireless Network.
Wireless Networks, January 2005.

[27] Kotz, D. and Henderson, T. CRAWDAD: A Community Resource for
Archiving Wireless Data at Dartmouth. IEEE Pervasive Computing, 2005.
[28] Kumar, U., Thakur, G. and Helmy, A. PROTECT: Proximity-based
Trust-advisor using Encounters for Mobile Societies. ACM IWCMC 2010.
[29] Lee, J. and Hou, J. C. Modeling steady-state and transient behaviors of
user mobility: formulation, analysis, and application ACM MobiHoc 2006
[30] Lelescu, D., Kozat, U. C., Jain, R. and Balakrishnan, M. Model T++: :
an empirical joint space-time registration model. ACMMobtHoc, 2006.
[31] McPherson, M., Smith-Lovin, L. and Cook, J. M. Birds of a Feather:
Homophily in Social Networks. Annual Review of Sociology, 2001.
[32] Mtibaa, A., May, M., Diot, C. and Ammar, M. PeopleRank: Social
Opportunistic Forwarding. IEEE INFOCOM, March 2010.
[33] Musolesi, M., Hui, P., Mascolo, C. and Crowcroft, J. Writing on the
Clean Slate: Implementing a Socially-Aware Protocol in Haggle. IEEE
Int'l Workshop on Autonomic and Opportunistic Communications, 2008.
[34] Musolesi, M. and Mascolo, C. A community based mobility model for
ad hoc network research. ACM REALMAN workshop, 2006.
[35] Nazir, F., Prendinger, H. and Seneviratne, A. Participatory mobile
social network simulation environment. MobiOpp '10.
[36] Newman, M. E. J. Modularity and community structure in networks.
Proceedings of the National Academy of Sciences, 103, 23, June 2006.
[37] Newman, M., Barabasi, A. and Watts, D. J. The Structure and
Dynamics ofNetworks: Princeton University Press, 2006.
[38] NOMADS Lab. MobiLib: Community-wide Library of Mobility and
Wireless Networks Measurements http://nile.cise.ufl.edu/MobiLib/
[39] Reddy, S., Estrin, D. and Srivastava, M., Recruitment Framework for
Participatory Sensing Data Collections", International Conference on
Pervasive Computing, May, 2010.
[40] Rhee, I., Shin, M., Hong, S., Lee, K. and Chong, S. On the Levy-Walk
Nature of Human Mobility, IEEE INFOCOM, pp. 924-932, 2008.
[41] Royer, E. M., Melliar-smith, P. M. and Moser, L. E. An Analysis of the
Optimum Node Density for Ad hoc Mobile Networks. IEEE International
Conference on Communications. ()., 857-861.
[42] Schutt, R. K. Investigating the Social World: The Process and Practice
ofResearch. Pine Forge Press, 2006.
[43] Shilton, K., Ramanathan, N., Reddy, S., Samanta, V., Burke, J. A.,
Estrin, D., Hansen, M. and Srivastava, M. B. Participatory Design of
Sensing Networks: Strengths and Challenges. Participatory Design
Conference, Bloomington, 2008.
[44] Spyropoulos, T., Psounis, K. and Raghavendra, C. S. Performance
analysis of mobility-assisted routing. ACMMobiHoc, 2006.
[45] Steinhaeuser, K. and Chawla, N. V. Community Detection in a Large
Real-World Social Network. Social Computing, Behavioral Modehng, and
Prediction, Springer, 2008.
[46] Thakur, G. S. Technical Report on Similarity in TVC Model,
http://www.cise.ufl.edu/techreport/simtvc.pdf, 2010.
[47] Wei-jen Hsu and Helmy, A. On Modeling User Associations in
Wireless LAN Traces on University Campuses. IEEE Int.l Workshop on
Wireless Network Measurements 2006.
[48] Yoneki, E., Hui, P. and Crowcroft, J. Visualizing community detection
in opportunistic networks. ACM CHANTS workshop, pp. 93-96, 2007.
[49] Zachary, W. W. An information flow model for conflict and fission in
small groups. Journal of Anthropological Research, pp. 452-473, 1977.


Following are the results generated from the TVC that shows TVC do

1. Location Visiting preferences
2. Periodical Reappearances.

Figure 4(a), with communities exits via TVC

As shown in the Figure 7-9 we can see that TVC Model captures
the basic characteristics of the location visiting preferences and
periodic reappearances. From there we can see that, despite TVC
is able to capture these characteristics, the explicit similarity is not
seen the TVC model unless we propose some important measure
that helps to understand the similarity distribution in some way.

. 0.6.


o 0.3




S Community model


Ran om action mndeI USC MT

0 1 2 3 4 5 6 7
Time gap (days)

Figure 10 Insight into the real distribution in the traces The figure shows the re-
appearance probabilities as calculated from the real traces

Figure 7 The location visiting preferences for the MIT Trace

USC Vs TVC Generated Re appearance plot

0 1 2 3 4 5 6 7
Time Gap in Days

Figure 8 The Periodic Re-appearances for the USC trace



0.20 Model-com lex

1 Model-simplified
S 0 1 2 3 4 5 6 7 8
Time gap (days)
Figure 9 Periodic Reappearances for the MIT Trace

Figure 11 Distribution of similarity for first two weeks for the USC Real trace

The pie chart in Figure-11, shows the distribution of similarity for
the user pairs of the USC campus. From the pie-chart it is more
clear that similarity is distributed for all different values and there
exists different types of user having varied similarities with other

14%[<1] 11%[>9] 15%, 11 14% [>91


8 9 10

University of Florida Home Page
© 2004 - 2010 University of Florida George A. Smathers Libraries.
All rights reserved.

Acceptable Use, Copyright, and Disclaimer Statement
Last updated October 10, 2010 - - mvs